Voronoi-based decision boundaries

This repository contains notebooks focused on decision boundary plotting for high-dimensional data. The main idea is to approximate decision boundaries in 2D by superimposing Voronoi tessellations onto the 2D-transformed data. To accomplish this, a 1-nearest-neighbour (1-NN) model is used to approximate the original model's predictive behaviour on the original (N-dimensional) data and apply the same technique on 2D-transformed data.

The plots below illustrate the Voronoi-based decision boundaries estimated for three different models trained on 6D data with 4 target classes. It can be seen that the model is able to make some sense of which model is overfitting and which is not.

Project structure

This project contains two Jupyter notebooks and a utility script. They are organised as follows.

.
├── Decision boundary plotting.ipynb    # notebook for decision boundary plotting
├── Voronoi diagram.ipynb               # notebook for Voronoi illustration
├── utils.py                            # utility script containing helper functions for plotting
└── README.md

References

• Visualizing multi-dimensional decision boundaries in 2D by Migut et al.
• Wikipedia article on Voronoi diagrams